Intra and Interdomain Circuit Provisioning Using the OSCARS Reservation System

Page created by Alberto Fowler
 
CONTINUE READING
Intra and Interdomain Circuit Provisioning Using the OSCARS Reservation System
Intra and Interdomain Circuit Provisioning Using
           the OSCARS Reservation System
     Chin Guok∗, David Robertson∗† , Mary Thompson†, Jason Lee† , Brian Tierney† and William Johnston∗†
                                                    ∗ Energy
                                                          Sciences Network
                                   Berkeley, California 94720; Email: {chin, wej}@es.net
                                  † Ernest Orlando Lawrence Berkeley National Laboratory
                      Berkeley, CA 94720; Email: {dwrobertson, mrthompson, jrlee, bltierney}@lbl.gov

  Abstract— With the advent of service sensitive applications          There are significant challenges to allowing users to sched-
such as remote controlled experiments, time constrained massive     ule high-performance network services on a production net-
data transfers, and video-conferencing, it has become apparent      work. Some of these challenges are: allowing only authorized
that there is a need for the setup of dynamically provisioned,
quality of service enabled virtual circuits. The ESnet On-Demand    users to create and manage high-performance services; pro-
Secure Circuits and Advance Reservation System (OSCARS) is          viding an easy-to-use interface for scheduling and managing
a prototype service enabling advance reservation of guaranteed      network resources; limiting the impact of high-performance
bandwidth secure virtual circuits.                                  traffic on other network traffic; coordinating quality of service
  OSCARS operates within the Energy Sciences Network                end-to-end across more than one autonomous network domain;
(ESnet), and has provisions for interoperation with other network
domains. ESnet is a high-speed network serving thousands of         and handling changes in network paths between the time a
Department of Energy scientists and collaborators worldwide.        reservation is scheduled and when it is claimed. OSCARS
  OSCARS utilizes the Web services model and standards to           currently addresses all but the last item, where work is still in
implement communication with the system and between domains,        progress.
and for authentication, authorization, and auditing (AAA). The         Due to the highly distributed nature of large scale science,
management and operation of end-to-end virtual circuits within
the network is done at the layer 3 network level. Multi-Protocol    the service framework for OSCARS is being developed in co-
Label Switching (MPLS) and the Resource Reservation Protocol        ordination with other community network provisioning efforts.
(RSVP) are used to create the virtual circuits or Label Switched    The closest coordination is with the Internet2 BRUW [23]
Paths (LSP’s). Quality of Service (QoS) is used to provide          system. A version of BRUW was used as the starting point
bandwidth guarantees.                                               for OSCARS, and now the two projects share a common code
  This paper describes our experience in implementing
OSCARS, collaborations with other bandwidth-reservation             base. Interoperability testing is on-going with Internet2 sites.
projects (including interdomain testing) and future work to be         OSCARS has been deployed within ESnet, which is a
done.                                                               nation-wide network that serves approximately 42 directly
                                                                    connected sites around the country. Internally, ESnet manages
                      I. I NTRODUCTION                              about 270 routers and systems throughout the network and
    Large-scale science is increasingly important as attention      its operations centers. The current ESnet architecture is that
turns to the study of the most complex, subtle, and elusive         of a high-bandwidth (10/2.5 Gb/s) backbone ring around the
natural phenomena. Such study is completely dependent on            country, with hubs at strategic locations. The sites, which are
world-wide collaborations of scientists and widely dispersed        mainly large Office of Science laboratories, are connected to
resources such as computing, data, and instruments.                 the hubs via metro rings at 2 x 10Gb/s speeds. OSCARS faces
    Over the past several years significant improvements have       the constraint of operation within this production network,
been made in the computing and communications infrastruc-           where 99.9+% reliability is critical.
ture necessary for support of these collaborations. Network            Having the ability to dynamically allocate capacity in the
bandwidths have increased, data transport protocols have im-        network exclusively to a scheduled service, to the exclusion
proved, and security issues have become better understood.          of normal priority traffic, introduces risks. Throughout the
However, for the network to fully enable such distributed sci-      design and implementation of OSCARS, security aspects were
ence, network communication must be delivered as a manage-          paramount. The impact of an abuse could be very large. A
able service to the distributed applications just as computing      denial of service attack could prevent reservations from being
is.                                                                 processed. If the service is compromised, an attacker could
    The goal of OSCARS is to manage and schedule high-              disable the wide area network.
impact network services associated with these collaborations.          This paper describes how OSCARS addresses the above
These services, which move multi-terabyte to multi-petabyte         constraints and risks while implementing a scheduling system.
datasets from experiments and simulations, and may include          Section II covers the OSCARS architecture, and Section III
high-end remote visualizations, cannot be provided cost-            describes the details of path setup and reservation handling.
effectively by best-effort service on a production network.         Section IV outlines some issues with cross-domain interoper-
vation details, modify and/or cancel one or more existing reser-
                                                                                   vations, and provide a summary of all current reservations.
                                                                                      To perform these functions, the BSS keeps information
                                                                                   about past, pending, and current reservations, and tracks the
                                                                                   current topology and state of the network. As part of schedul-
                                                                                   ing a reservation request, the BSS must determine whether
                                                                                   the requested bandwidth will over-subscribe any of the links
                                                                                   in the path to be set up within the network.
                       Fig. 1.   OSCARS Architecture.
                                                                                      The PSS is responsible for setting up and tearing down the
                                                                                   on-demand bandwidth paths. This is accomplished by making
ability, and covers other reservation systems and collaborative                    the necessary configuration changes in the routers to create
efforts, as well as an example of an interoperability test                         or destroy a Label Switched Path (LSP) at the time indicated
between OSCARS and BRUW. Section V covers security,                                by the BSS. The authentication and authorization method for
including system, interdomain, and authentication, authoriza-                      the PSS is internal to the ESnet network and is specific to the
tion and auditing issues. The final section touches briefly                        router platform (currently Juniper or Cisco) being configured.
on the future work that needs to be performed to handle a                          It is therefore distinct from the AAAS used by the BSS.
dynamically changing network (which may invalidate exist-                          B. Implementation
ing reservations), and to set up end-to-end circuits between
                                                                                      Web services standards are used wherever possible. SOAP
domains in a secure, standards-based fashion.
                                                                                   messages are used for communications between clients and
                          II. A RCHITECTURE                                        the RM, and the W3C Web Services Definition Language
                                                                                   (WSDL) is used for the service description.
  The intent of OSCARS is to create a service for dynamic                             The resource manager is implemented as an Apache Web
QoS path establishment that is simple for users to use, and                        server configured for mod perl, a SOAP server, two databases
easy to administer. The only task required of a user is to                         within a MySQL server, and a set of Perl packages that
make a bandwidth reservation. Reservation can be made either                       implement the AAAS, BSS, and PSS. A mod perl script on
for immediate use or in advance for either one-time use or                         the Web server is used to accept browser requests sent via
persistent use, e.g. for the same time everyday. The user does                     Asynchronous Javascript + XML [13] techniques. Parameters
not have to configure an alternate routing path, nor mark                          from the Ajax request are then placed into a SOAP request,
the packets in any way. All necessary mechanisms needed                            and forwarded to the SOAP server. URL’s indicating an
to provide the user with a guaranteed bandwidth path are                           explicit SOAP request conforming to the service description
coordinated by a Reservation Manager (RM) and managed                              are proxied by the Web server directly to the SOAP server.
by the routers in the network.                                                     Note that these latter requests require that the SOAP message
                                                                                   be signed. Security issues are discussed in more detail in the
A. Components
                                                                                   section on security below.
   The RM is comprised of three components: the Authenti-                             The SOAP server makes database requests as necessary,
cation, Authorization, and Auditing Subsystem (AAAS), the                          depending on the method called. One database contains per-
Bandwidth Scheduler Subsystem (BSS), and the Path Setup                            sistent information related to methods handling reservations
Subsystem (PSS) (Figure 1). All persistent information is                          and AAA, and the other contains a representation of the
stored in a database. The RM provides simple Web forms                             local network topology. The latter is used primarily during
for creating and managing reservations, setting authorization                      reservation setup.
policy and other administrative tasks (the Web-based user
interface in the figure). It also supports an API using the                                        III. PATHS AND R ESERVATIONS
W3C SOAP1 messaging protocol to support programmable                               A. Path Setup
reservation management and requests from other network                               The procedure of a typical path setup is as follows:
service providers.                                                                   1) A user submits a request to the RM (using either an API
   The AAAS is responsible for authenticating and authorizing                           or an optional Web front-end) to schedule an end-to-end
all external requests, for logging request information, and                             path (e.g. between an experiment and computing cluster)
sending notifications to users and administrators of the results                        specifying start and end times, bandwidth requirements,
of calls made to the RM. It also handles a number of internal                           the source host that will be used to provide an applica-
requests related to management of users and resources.                                  tion access to the path, and the destination host.
   The BSS is responsible for scheduling reservations. It                            2) User parameters are validated by the RM, to ensure
handles requests to schedule bandwidth reservations, list reser-                        that they have the correct format, and to prevent SQL
  1 Bold-faced text indicates a specification that is only available online (see
                                                                                        injection attacks.
http://www.w3.org for W3C and http://oasis-open.org for                              3) Using the source and destination host information sub-
OASIS).                                                                                 mitted by the user, a traceroute is executed to determine
the path within ESnet that the MPLS LSP will traverse,        vested from the Border Gateway Protocol (BGP) on the egress
       as well as the ingress and egress border routers that will    PE router is used to determine the next Autonomous System
       originate and terminate the LSP.                              (AS) that the request should be forwarded to. The AS number
  4)   This information is stored by the BSS in a database, and      is checked against a list of known administrative domains that
       a script periodically checks to see if the PSS needs to       have reservation systems that are cooperating with OSCARS.
       be contacted, either to create or tear down the circuit.      If a match is found, the request is forwarded to the downstream
  5)   At the requested start time, the PSS configures the ESnet     AS.
       provider edge (PE) router (at the start end of the path)
       to create an LSP with the specified bandwidth.                C. Advanced Reservations
  6)   Each router along the route receives the path setup              To support advanced reservations, RSVP-TE [26] infor-
       request via the Reservation Resource Protocol (RSVP)          mation on the network’s current provisioned bandwidth, if
       [6] and commits bandwidth (if available) creating an          available, must be supplemented with a mechanism that will
       end-to-end LSP. The RM is notified by RSVP if the             permit a view of the network’s provisioned bandwidth in a
       end-to-end path cannot be established.                        slice of time in the future. In OSCARS, advance reservations
  7)   Packets from the source (e.g. experiment) are routed          are handled in a slot based manner. This equates to “first come
       through the site’s LAN production path to ESnet’s PE          first served” for bandwidth across any particular link at any
       router. On entering the PE router, these packets are          moment.
       identified and filtered using flow specification param-          As each reservation is requested in OSCARS, the end-to-
       eters (e.g. source/destination IP address/port numbers)       end path is computed for that reservation. Once the entire
       and policed at the specified bandwidth. The packets are       path through all the routers controlled by OSCARS has been
       then injected into the LSP and switched (using MPLS)          computed, each link in the path is checked for available
       through the network to its destination (e.g. computing        bandwidth. To check the bandwidth of a link, all outstanding
       cluster).                                                     reservations for that link during the time of the proposed
  8)   A notification of the success or failure of LSP setup is      reservation are queried from the data base. Then all the
       passed back to the RM so that the user can be notified        reserved bandwidth amounts are calculated and compared to
       and the event logged for auditing purposes.                   the actual capacity of the link. If the requested amount of
  9)   At the requested end time, the PSS tears down the LSP.        bandwidth plus all outstanding reservations is more then the
                                                                     allocated amount of bandwidth available for reservations on
B. Path Discovery                                                    that link (in this case 50%), then the reservation fails. Only
   There are two scenarios for creating a path in OSCARS.            if there is enough bandwidth available on all links is the
One is where the reservation request does not contain any            reservation committed into the reservation system.
connectivity information outside of the source and destination          In the future, in the case where the capacity of a link
(IP addresses). The other is where a request contains additional     changes (e.g. a link upgrade or failure), all outstanding reserva-
routing information such as the ingress and/or egress PE             tions that involve the use of that link will be queried from the
routers within the OSCARS administrative domain.                     system and recomputed. This will be done to ensure that there
   In the scenario where an ingress PE router is not explicitly      is adequate bandwidth available on the link when it comes
communicated, OSCARS does a traceroute (from the core of             time to provision the reservation.
the network) towards the source IP address of the traffic. As
the traceroute progresses, each router in the trace is checked       D. Provisioning and Policing
to verify if it is within the administrative control of OSCARS.         With reservations (current and advanced) managed by the
As soon as OSCARS encounters a router that is outside of its         OSCARS database, provisioning and policing are the next
administrative domain, OSCARS marks the last router (within          steps in permitting the reservation to be claimed. These steps
its administrative control) as the ingress PE router.                are necessary in order to facilitate bandwidth guarantees to
   In the scenario where the egress PE router is not contained       enforce reservation and usage limits. In OSCARS, RSVP
in the reservation request, the destination IP address is used       is used as the provisioning mechanism to instantiate and
as the target of the traceroute (sourced from the ingress            manage active reservations. However, RSVP does not police
PE router). Using the same method outlined in the previous           the usage limits of the reservations. To ensure that bandwidth
paragraph, the egress PE router can be determined.                   resources are not over-subscribed, QoS is carefully configured
   With the ingress and egress PE routers identified, the path       to provision queues within the network core.
(or route) between the two can be trivially determined.                 Within ESnet, traffic utilizing the OSCARS service is clas-
   The need for OSCARS to support reservations with explicit         sified into a Class-of-Service distinct from all other traffic and
ingress and egress PE routers is to facilitate traffic engineering   isolated into a separate queue by itself. The size and transmit
for sites or networks that have more then one peering connec-        rate of this queue is configured to match the RSVP bandwidth
tion with ESnet.                                                     limits on each interface, e.g. if the RSVP bandwidth limit
   In the event that the virtual circuit extends beyond              on an interface is 50%, the OSCARS queue depth and service
OSCARS’ administrative domain, routing information har-              rate is also set at 50%. This ensures that the RSVP provisioned
bandwidth will translate to available network bandwidth within      funded CHEETAH [27] and DRAGON [30] projects, Inter-
the network core.                                                   net2’s BRUW [23] and HOPI [3] projects, CANARIE’s UCLP
   With all of OSCARS traffic using the same Class-of-Service       [29] project, and GEANT’s BoD (GN2-JRA3) [25] and AMPS
queue within the network core, it is vital to ensure that           (SA3) [21] activities.
the bandwidth usage of each individual RSVP reservation is             With the exception of Lambda Station and TeraPaths, which
strictly adhered to. This prevents the aggregate traffic from       address the local network last-mile issues, all the other projects
overrunning the queue dedicated to the OSCARS service. To           mentioned above address issues in the wide-area network.
do this, each flow utilizing the OSCARS service is policed             All of these projects, as well as OSCARS, are based on a
individually according to the reservation bandwidth request.        Web services interface to reserve and configure a path across
This policing is done at the ingress point to ESnet.                the network. However, they have slightly different ways of
                                                                    handling reservations and AAA issues.
              IV. I NTERDOMAIN R ESERVATIONS                           Dragon uses OSPF-TE [19] for intradomain routing, and
                                                                    a component called the Network Aware Resource Broker
   Guaranteed bandwidth paths are most effective when the
reservation spans end-to-end. This however, introduces the          (NARB) for interdomain routing. Dragon plans to use the
                                                                    Common Open Policy Service (COPS) [11] protocol for
complexity of extending virtual circuits beyond the scope of a
                                                                    support of policy provisioning (COPS-PR) [9].
single administrative domain to multiple domains. To facilitate
this, neighboring domains must agree on several levels, mainly,        Canarie, Canada’s advanced Internet development organi-
the management plane, control plane, and data plane:                zation, has been working on a project called User Controlled
                                                                    LightPath (UCLP). UCLP allows end-users to create their own
  •   The management plane dictates policies and procedures         static independent IP network as a subset of a larger optical
      for authentication, authorization, and usage. This es-        network and to have total control over their share of network.
      sentially amounts to a Service Level Agreement (SLA)             The University of Amsterdam’s Advanced Internet Research
      between peer networks. In almost all cases, the usage con-    group has published a number of papers describing both the
      dition outlined within an SLA determines the maximum          networking and the AAA issues for such a system, including
      aggregate limit. This implies that individual bandwidth       [14] [15] [10] . They are using the IETF AAA Framework [28],
      requests are managed by the reservation system of the         and use the OASIS eXtensible Access Control Markup Lan-
      originating AS and not propagated independently to the        guage (XACML) Version 2 to describe policy. They have also
      transit AS’s (i.e. the transit AS will see the request as     defined a Network Description Language, which is a RDF-
      coming from the originating AS and not the individual         based method to describe networks, to facilitate interdomain
      making the request).                                          interoperability [16].
  •   The control plane dictates the way control messages, such        In the OSCARS, BRUW, TeraPaths, and AMPS projects, IP
      as setup and teardown requests, are exchanged between         connectivity (layer 3) is used as the data plane exchange. This
      the networks, e.g. RSVP signaling. At this point in           facilitates interoperability trials with no additional network
      time, interdomain interoperability efforts do not permit      connection needed outside of the production peering exchange.
      the end-to-end signaling of LSP’s via the control plane          It should be noted that OSCARS and BRUW now share
      (i.e. interdomain exchange of RSVP messages). This            a common code base, but are configured differently due to
      is because there is no vendor implementation that can         differing methods of network administration and user authen-
      enforce complex SLA requirements of the various ad-           tication.
      ministrative domains. As such, end-to-end virtual circuits       The other projects mentioned use the optical network layer
      are comprised of intradomain LSP’s stitched together at       (i.e. layer 1), creating lightpaths. The last section of this paper
      agreed interconnection points.                                points out the necessity of future work to bridge projects using
  •   The data plane handles how user traffic is forwarded from     different layers.
      one network to another network, e.g. IP packets, Ethernet
      VLAN packets, etc. [2]. This is one of the fundamental        B. Interdomain Path Setup
      issues that must be resolved in order for an interdomain         One of the more complex examples of using the OSCARS
      end-to-end virtual circuit to be successful. Complications    reservation system involves the setting up of a virtual circuit
      arise when peering RM’s provision virtual circuits at         between two hosts that span several administrative domains.
      different network layers (e.g. GMPLS LSP, MPLS LSP).          For example, imagine setting up a virtual circuit between
      The solution to bridging the data planes is part of ongoing   Host A and Host B, where Host A is controlled by a remote
      collaborative efforts.                                        reservation system RM A in ISP A, and Host B is part of
                                                                    the local OSCARS reservation system in ISP B (see Figure
A. Related Work                                                     2). The routed path from ISP B to ISP A transits ISP X, but
  There are several implementations of network resource             the reverse path from ISP A to ISP B is via ISP Y. In order
management and service provisioning systems in existence            for an OSCARS’ user to make an interdomain virtual circuit
today. These include the DOE funded Lambda Station [4],             reservation request from Host A to Host B, the following must
TeraPaths [5] and UltraScience Net [22] projects, the NSF           occur.
Test           Test        Guaranteed    Throughput
                                                                               Protocol      Parameters     Bandwidth      Achieved
                                                                                UDP         Throughput          No          30Mb/s
                                                                                            Set: 30Mb/s    Yes (25Mb/s)    24.6Mb/s
                                                                                TCP        TCP Window           No        158.0 Mb/s
                                                                                             Size: 1MB     Yes (25Mb/s)    14.7Mb/s
                                                                                          Latency: 50 ms

                                                                                                     TABLE I
                                                                               IP ERF RESULTS OF GUARANTEED BANDWIDTH PATHS .

                                                                      and an ESnet Performance Center [12] host in Sunnyvale, CA.
                                                                      The path consisted of two unidirectional MPLS LSP’s, one
                                                                      in Abilene, and the other in ESnet, stitched together at the
                                                                      Abilene-ESnet peering point in Chicago, IL.
                                                                         Bandwidth tests using IPerf [18] (see Table I) revealed
                  Fig. 2.   Interdomain Path Setup.                   predictable results except for the guaranteed bandwidth TCP
                                                                      transfer.
                                                                         The guaranteed bandwidth (25Mb/s) TCP transfer should
   1) On receiving the request from the user, OSCARS com-
                                                                      have yielded a throughput closer to 25Mb/s. On further inves-
       putes the virtual circuit path and determines the down-
                                                                      tigation, it was determined that the discrepancy between the
       stream AS (ISP X).
                                                                      policing bandwidth and the achieved bandwidth was likely
   2) The request is then encapsulated in a message forwarded
                                                                      the result of two things, first, the lack of traffic shaping at the
       across the network (ISP X) towards Host A, crossing
                                                                      source end, and second, Juniper’s policing function. Similar
       all intervening reservations systems (RM X), until it
                                                                      results have been documented by others [20].
       reaches the last reservation system (RM A) that has
       administrative control over the network (ISP A) that                                      V. S ECURITY
       Host A is attached to.                                            Since OSCARS is being deployed on the ESnet produc-
   3) The remote reservation system (RM A) then computes              tion network, security was an absolute requirement from the
       the path of the virtual circuit, and initiates the bandwidth   beginning. Good security needs mechanisms that are easy to
       reservation requests from Host A towards Host B (via           understand, install, use, and administer. It is very important
       ISP Y). This can be especially complex when the path           that there are no unintended consequences of authorization
       back (from Host B to A) is asymmetric and traverses            policy decisions.
       AS’s (e.g. ISP Y) that were not traversed on the for-             The following section details steps taken to secure the
       ward path, causing the local OSCARS to see the path            machines and servers running OSCARS, and the remaining
       originating from a different AS than it originally sent        security sections cover AAA.
       the request to.
   To facilitate interdomain virtual circuit setup, a WSI-BP          A. System
[1] compliant WSDL specification for the network-network                 An Apache2 Web server on an open ESnet machine is
interface has been written following the model of GEANT’s             used to forward all requests to the RM Web server, which
Advance Multi-domain Provisioning System (AMPS) [21] .                runs on a machine behind a firewall. This forwarding process
This interface is being tested with the TeraPaths [5] project.        is transparent to the end user, and hides the location of the
Having a WSDL specification allows reservation systems to             internal server.
communicate with one another in a well defined syntax.                   The internal Web server only accepts https connections from
While the OSCARS interface is similar to the one specified            the open machine. The RM SOAP server only accepts requests
by AMPS, these are both quite different from the TeraPaths            from the Web server on the internal machine or digitally signed
interdomain WSDL. One of the next challenges in automating            SOAP messages encapsulated in https messages forwarded
trans-domain circuit setup is to define a standard request for        from the open Web server. Database server processes run as
interdomain reservations. It would then be up to the individual       an unprivileged user without a login shell. The database server
reservation system to transform the standardized messages to          only accepts requests from the SOAP server.
internal calls to reserve and provision the virtual circuit.
                                                                      B. Authentication
C. Interoperability Tests                                                OSCARS authenticates the sender of all requests that it
  In April 2006, an interdomain guaranteed bandwidth path             receives. The Web based interface and the SOAP server API
between Abilene and ESnet was dynamically negotiated and              use the authentication mechanism that is most natural for them.
configured by the BRUW and OSCARS systems respectively.               The Web server requires a username and password for au-
The unidirectional 25Mb/s guaranteed bandwidth path was               thentication on the first access during which it creates session
configured between an Internet2 test host in Indianapolis, IN         information for the user and a 8-hour cookie referencing this
information. This cookie is used on subsequent connections.          that facilitate the expression of complex access policy, they
All communication with the Web server takes place over               typically require the use of a parsing engine and interpreter to
encrypted https in order to protect against the stealing of          evaluate a request for action. When even moderately complex
passwords or cookies. The SOAP API distinguishes between             policies are used, it becomes difficult to tell at a glance exactly
requests coming from the Web server on the local host, which         what access is allowed. With the simple database policy, it
it assumes have been authenticated as just described, and            is easy to query exactly who has access to a resource and
requests coming from the open Web server. The latter requests        what resources a particular user or AS has. Thus an ESnet
must be digitally signed messages signed by the originator of        system administrator who is not directly part of the OSCARS
the message. The OASIS WS-Security X509 profile rules and            implementation team can use familiar tools to check on (or
syntax are used for the signing. The AAAS verifies the signa-        modify) who has rights to make reservations or control routing
ture and the signing certificate to authenticate the user. It then   on “his” network.
uses the subject name from the signing certificate to identify          Within OSCARS, access is controlled for the creation and
the user. Because there is a Web server on the open network          management of reservations, users and domains. The permis-
interposed between the requester and the SOAP server behind          sions that can be granted are viewing or managing (modifying)
the firewall, digitally signed messages are needed to do end-        a resource, creating reservations, and specifying routing.
to-end client authentication and to support proxy certificates as       As long as there are not too many resources, permissions, or
a single-signon mechanism. Both the username/password and            users, this method provides an easily managed and understood
certificate authentication methods use the database user table       access control scheme. For practical purposes, having to scale
to determine if the request is coming from a legitimate user.        to larger numbers is not an issue. The maximum number of
This table contains a mapping of the OSCARS user name,               users who will be able to use the OSCARS system is small,
password, subject name from the certificate and the certificate      consisting of users at ESnet sites responsible for moving large
itself, as well as other information about the user.                 amounts of information. Users from other domains will be
   Requests for or about interdomain reservations are authen-        authorized by their own RM.
ticated in the originating domain on the basis of an individual
user, and in the subsequent domains on the basis of the              D. Auditing
RM in the adjacent domain. This approach follows the AAA                At this point, the OSCARS server logs all significant activity
Authorization model defined by the IETF Networking Group.            such as creating or canceling reservations. In addition a list of
[28]. In this model users are authenticated and authorized for       all reservations is kept and can be read via a SOAP request.
actions in their home domain and interdomain authorization           As was mentioned above, in interdomain requests, the name
depends on SLA’s between domains (AS’s) and the assurance            of the originating user is passed to the next domain where it
that a request is coming from a trusted server in a trusted          can be used for either authorization or auditing.
domain. Normally all requests forwarded between domains
will be SOAP messages signed by the RM. The OSCARS                             VI. C ONCLUSIONS       AND   F UTURE W ORK
RM has a list of the cooperating RM certificates as well as a          OSCARS is one example of a system that will become
list of permissions for those AS’s. In effect a service level        increasingly necessary as experiments such as the Large
agreement gets implemented in the user and authorization             Hadron Collider become operational. It allows users to easily
tables in the database. The message forwarded adds the name          schedule in advance the network bandwidth necessary for
of the originating user, in case other domains wish to use           their experiment or simulation. Since it provides the ability
that information for authorization or auditing. Currently, at        to change router configurations in a production network,
the time of provisioning no further authentication is done.          maintaining security is an integral part of its operation.
Provisioning is triggered by the time of the reservation. Once         A number of issues need to be addressed before such
the provisioning has been completed, any traffic coming from         systems become production level in complex network envi-
the specified ingress router is able to use the higher class of      ronments where many autonomous domains may be involved,
bandwidth.                                                           and where network topologies may be constantly changing.
C. Authorization                                                     A. Topology Changes
   User’s authorizations to operate on resources are stored as          A key consideration, when running OSCARS as a produc-
relations in database tables. After a user has been authenti-        tion service, is the ability to recover from both scheduled and
cated, their request is checked by the RM to see if the user’s       unscheduled network outages or changes. This is particularly
authorizations are sufficient.                                       complex when dealing with bandwidth reservations made for
   This implementation is similar to the ROAM authorization          a future date. For example, in the event of an unscheduled
service of the FusionGrid [7]. This approach allows the use of       network outage, future reservations committed on the affected
standard database commands to define resources and permis-           links must be recalculated. This can be further complicated
sions and to manage and check authorizations. While there are        if the outage period is unknown. The converse is also true.
many “policy languages” (e.g. S-expressions [24], the OASIS          If a new link were to be added or upgraded, increasing
Security Assertion Markup Language (SAML), and XACML)                the bandwidth allocation for future reservations creates an
inconsistent view with the current state of the network. To                                    ACKNOWLEDGMENT
deal with this in OSCARS, a polling mechanism is being                  This work was supported by the Director, Office of Science.
developed that constantly compares the current state of the           Office of Advanced Scientific Computing Research. Mathe-
network to the state that is reflected in the OSCARS topology         matical, Information, and Computational Sciences Division
database. If there are differences that affect the characteristics    under U.S. Department of Energy Contract No. DE-AC02-
of a link, the new topology is pushed into the OSCARS                 05CH11231. This is LBNL report number LBNL-60373.
database and all outstanding reservations that use that link are        The authors would like to thank Bob Riddle and Andrew
queried from the database. This is possible since the complete        Lake of Internet2, both for help in the incorporation of BRUW,
path of all reservations is kept in the database along with the       and for work on interoperability tests.
reservation. Then all the reservations are recomputed in the
order they were placed into the OSCARS system, to ensure                                            R EFERENCES
that the requested resources are still available. In the event that   [1] K. Ballinger, D. Ehnebusk, M. Gudgin, M. Nottingham, and
the necessary resources are unavailable for a reservation that            P. Yendluri, “Basic profile version 1.0,” [Online document] Apr.
has been entered into OSCARS, the reservation’s state will                2004,      Available    at:    http://www.ws-i.org/Profiles/
                                                                          BasicProfile-1.0-2004-04-16.html
be changed to unavailable and a notice will be sent to the            [2] E. Bell, A. Smith, P. Langille, A. Rijhsinghani, and K. McLoghrie,
administrators informing them of the problem.                             “Definitions of managed objects for bridges with traffic classes, multicast
                                                                          filtering and virtual LAN extensions,” IETF RFC 2674, Aug. 1999.
                                                                      [3] H. Boyles, “Recent Results from Internet2’s Hybrid Optical and Packet
B. Network-Network Interface                                              Infrastructure Project (HOPI),” Presented at TERENA Networking Con-
                                                                          ference, Rhodes, Greece, 2004.
  With the objective of interoperability between the different        [4] A. Bobyshev, M. Crawford, P. DeMar, V. Grigaliuna, M. Grigoriev,
                                                                          A. Moibenko, D. Petravick, R. Rechenmacher, H. Newman, J.
networks, comes the need for defining standard interfaces (i.e.           Bunn, F. van Lingen, D. Nae, S. Ravot, C. Steenberg, X. Su,
Network-to-Network Interface). This is to facilitate the shar-            M. Thomas, Y. Xia “Lambda Station: Production Applications
ing of network state and request information in quantifiable              Exploiting Advanced Networks in Data Intensive High Energy Physics,”
                                                                          in        Computing in High Energy and Nuclear Physics (CHEP) 2006,
characteristics that are common to collaborating networks.                Mumbai, India, Feb. 13-17, 2006.
This could include properties such as connectivity (topology),        [5] S. Bradley, F. Burstein, L. Cottrell, B. Gibbard, D. Ka-
bandwidth, latency, and jitter. There are several documents               tramatos, Y. Li, S. McKee, R. Pope-scu, D. Stampf,
                                                                          and D. Yu, “TeraPaths: A QoS-enabled collaborative data
that have been published by the different projects as well                sharing infra-structure for peta-scale computing research,” in
as organizations related to this work (e.g.: GEANT [25],                  Computing in High Energy and Nuclear Physics (CHEP) 2006,
the University of Amsterdam [16], Canarie’s UCLP project                  Mumbai, India, Feb. 13-17, 2006.
                                                                      [6] R. Branden, L. Zhang, S. Berson, S. Herzog, and S. Jamin, “Resource
[8], and the IETF CCAMP Working Group [17]). However,                     ReSerVation Protocol (RSVP) – Version 1 Functional Specification,”
due to the heterogeneity of network implementations and                   IETF RFC 2205, Sept. 1997.
deployments, generating a single framework to quantify all            [7] J.R. Burruss, T.W. Fredian, and M.R. Thompson, “Security on the U.S.
networks is challenging.                                                  FusionGrid,” to appear in Fusion Engineering and Design, Elsevier, Fall
                                                                          2006.
                                                                      [8] “Canarie interoperability work,” [Online document] Available at:
                                                                          http://grid2.canarie.ca/wiki/index.php/
C. Hybrid Data-Planes                                                 [9] K. Chan et al., “COPS usage for policy provisioning (COPS-PR),” IETF
                                                                          RFC 3084, Mar. 2001.
   With the emergence of numerous reservation systems, it is          [10] Y. Demchenko, L. Gommans, C. de Laat, A. Tok-
becoming evident that there is a need to bridge these systems             makoff, and R. van Buren, “Policy based access control
which provision virtual circuits at different network layers.             in      dynamic      Grid-based    collaborative     environment,”       in
                                                                          International Symposium on Collaborative Technologies and Systems,
For example, OSCARS and BRUW provision MPLS LSP’s                         pp. 64-73, 2006.
over an IP (layer 3) shared network, whereas DRAGON and               [11] D. Durham et al., “The COPS (Common Open Policy Service Protocol),”
CHEETAH use GMPLS to set up lightpaths over a lambda                      IETF RFC 2748, Jan. 2000.
                                                                      [12] “ESnet performance center,” [Online document] Available at:
switched (layer 1) network. The challenge here is to bridge               https://performance.es.net/
the two systems such that an end-to-end connection appears            [13] J.J. Garrett, “Ajax: A new approach to Web applications,” [Online
to be seamless to the end-user.                                           document] Feb. 2005, Available at: http://www.adaptivepath.
                                                                          com/publications/essays/archives/000385.php
   In addition, numerous complications arise when reservation         [14] L. Gommans, C. de Laat, and R. Meijer, “Token based path authorization
systems managing disparate data planes attempt to exchange                at interconnection points between hybrid networks and a lambda grid,”
connectivity information. First, there is a need to translate             in Proceedings of IEEE GRIDNETS 2005.
                                                                      [15] L. Gommans, B. van Oudenaarde, F. Dijkstra, C. de Laat, T. La-
or map connectivity information such that it is usable by                 vian, I. Monga, A. Taal, F. Travostino, and A. Wan, “Applica-
the reservation system receiving the information (e.g. layer              tions drive secure lightpath creation across heterogeneous domains,”
1 connectivity information is meaningless to a layer 3 reser-             IEEE Communications Magazine, vol. 44, no. 3, 2006.
                                                                      [16] J.J. van der Ham, F. Dijkstra, F. Travostino, H.M.A. Andree, and
vation system unless it is associated with an IP address). This           C.T.A.M. de Laat, “Using RDF to describe networks”, iGrid 2005 special
is the approach explored by DRAGON. Second, there needs                   issue, Future Generation Computer Systems, vol. 22, no. 8, pp. 862-867,
to be a mechanism to redistribute multi-layer connectivity                2006.
                                                                      [17] “IETF CCAMP Working Group,” [Online document] Available at:
information. Within the IP layer, this is done via BGP. At                http://www.ietf.org/html.charters/ccamp-charter.
lower network layers, no such comparable protocols exist.                 html
[18] “iperf,” [Online document] Available at:
    http://dast.nlanr.net/Projects/Iperf/
[19] K. Ishiguro, A. Davey, and A. Lindem (Ed.), “Traffic engineering
    extensions to OSPF version 3”, IETF Internet Draft, draft-ietf-ospf-
    ospfv3-traffic-07.txt, Apr. 2006.
[20] E. Martelli, “Protect production traffic against aggressive streams using
    Juniper routers,” [Online document] Aug. 2003, Available at: http:
    //emartell.web.cern.ch/emartell/done/datatag/
    juniper rate limit/traffic protection with juniper.
    html
[21] A. Patil, “Advance multi-domain provisioning system,” Presented at
    TERENA Networking Conference, Catania, Italy, 2006.
[22] N.S.V. Rao, W.R. Wing, S.M. Carter, and Q. Wu, “Ultra-
    Science Net: Network testbed for large-scale science applications,”
    IEEE Communications Magazine, vol. 43, no. 11, pp. S12-17, 2005.
[23] B. Riddle, “BRUW: A bandwidth reservation system to support end-user
    work”, Presented at TERENA Networking Conference, Poznan, Poland,
    2005.
[24] R. Rivest, “S-expressions”, Network Working Group, Internet Draft
    [Online document] May 1997, Available at
    http://theory.lcs.mit.edu/rivest/sexp.txt
[25] A. Sevasti, “GN2-JRA3: A multi-domain bandwidth on demand service
    for the NREN community,” Presented at TERENA Networking Confer-
    ence, Catania, Italy, 2006.
[26] G. Swallow, J. Drake, H. Ishimatsu, and Y. Rekhter. “Generalized
    Multiprotocol Label Switching (GMPLS) User-Network Interface (UNI):
    Resource ReserVation Protocol-Traffic Engineering (RSVP-TE) Support
    for the Overlay Model”, IETF RFC 4208, Oct. 2005.
[27] M. Veeraraghavan, X. Zheng, H. Lee, M. Gardner, and W. Feng, “CHEE-
    TAH: Circuit-switched high-speed end-to-end transport architecture,” in
    Proceedings OptiComm 2003, Dallas, TX, Oct. 13-17, 2003.
[28] J. Vollbrecht, P. Calhoun, S. Farrell, L. Gommans, G. Gross,
    B. de Bruijn, C. de Laat, M. Holdrege, and D. Spence, “AAA autho-
    rization framework,” IETF RFC 2904, Aug. 2000.
[29] J. Wu, M. Savoie, S. Campbell, H. Zhang, G.V. Bochmann, and
    B. St. Arnaud, “Customer-managed end-to-end lightpath provisioning,”
    International Journal of Network Management, vol. 15, no. 5, pp. 349-
    362, 2005.
[30] X. Yang, T. Lehman, C. Tracy, J. Sobieski, S. Gong, P. Torab,
    and B. Jabbari, “Policy-based resource management and service provi-
    sioning in GMPLS networks,” in First IEEE Workshop on “Adaptive
    Policy-based Management in Network Management and Control”, at
    IEEE INFOCOM 2006, Barcelona, Spain, Apr. 2006.
You can also read